Header logo is


2007


no image
Learning with Transformation Invariant Kernels

Walder, C., Chapelle, O.

(165), Max Planck Institute for Biological Cybernetics, Tübingen, Germany, September 2007 (techreport)

Abstract
Abstract. This paper considers kernels invariant to translation, rotation and dilation. We show that no non-trivial positive definite (p.d.) kernels exist which are radial and dilation invariant, only conditionally positive definite (c.p.d.) ones. Accordingly, we discuss the c.p.d. case and provide some novel analysis, including an elementary derivation of a c.p.d. representer theorem. On the practical side, we give a support vector machine (s.v.m.) algorithm for arbitrary c.p.d. kernels. For the thin-plate kernel this leads to a classifier with only one parameter (the amount of regularisation), which we demonstrate to be as effective as an s.v.m. with the Gaussian kernel, even though the Gaussian involves a second parameter (the length scale).

ei

PDF [BibTex]

2007


PDF [BibTex]


no image
Scalable Semidefinite Programming using Convex Perturbations

Kulis, B., Sra, S., Jegelka, S.

(TR-07-47), University of Texas, Austin, TX, USA, September 2007 (techreport)

Abstract
Several important machine learning problems can be modeled and solved via semidefinite programs. Often, researchers invoke off-the-shelf software for the associated optimization, which can be inappropriate for many applications due to computational and storage requirements. In this paper, we introduce the use of convex perturbations for semidefinite programs (SDPs). Using a particular perturbation function, we arrive at an algorithm for SDPs that has several advantages over existing techniques: a) it is simple, requiring only a few lines of MATLAB, b) it is a first-order method which makes it scalable, c) it can easily exploit the structure of a particular SDP to gain efficiency (e.g., when the constraint matrices are low-rank). We demonstrate on several machine learning applications that the proposed algorithm is effective in finding fast approximations to large-scale SDPs.

ei

PDF [BibTex]

PDF [BibTex]


no image
Sparse Multiscale Gaussian Process Regression

Walder, C., Kim, K., Schölkopf, B.

(162), Max Planck Institute for Biological Cybernetics, Tübingen, Germany, August 2007 (techreport)

Abstract
Most existing sparse Gaussian process (g.p.) models seek computational advantages by basing their computations on a set of m basis functions that are the covariance function of the g.p. with one of its two inputs fixed. We generalise this for the case of Gaussian covariance function, by basing our computations on m Gaussian basis functions with arbitrary diagonal covariance matrices (or length scales). For a fixed number of basis functions and any given criteria, this additional flexibility permits approximations no worse and typically better than was previously possible. Although we focus on g.p. regression, the central idea is applicable to all kernel based algorithms, such as the support vector machine. We perform gradient based optimisation of the marginal likelihood, which costs O(m2n) time where n is the number of data points, and compare the method to various other sparse g.p. methods. Our approach outperforms the other methods, particularly for the case of very few basis functions, i.e. a very high sparsity ratio.

ei

PDF [BibTex]

PDF [BibTex]


no image
Efficient Subwindow Search for Object Localization

Blaschko, M., Hofmann, T., Lampert, C.

(164), Max Planck Institute for Biological Cybernetics, Tübingen, Germany, August 2007 (techreport)

Abstract
Recent years have seen huge advances in object recognition from images. Recognition rates beyond 95% are the rule rather than the exception on many datasets. However, most state-of-the-art methods can only decide if an object is present or not. They are not able to provide information on the object location or extent within in the image. We report on a simple yet powerful scheme that extends many existing recognition methods to also perform localization of object bounding boxes. This is achieved by maximizing the classification score over all possible subrectangles in the image. Despite the impression that this would be computationally intractable, we show that in many situations efficient algorithms exist which solve a generalized maximum subrectangle problem. We show how our method is applicable to a variety object detection frameworks and demonstrate its performance by applying it to the popular bag of visual words model, achieving competitive results on the PASCAL VOC 2006 dataset.

ei

PDF [BibTex]

PDF [BibTex]


no image
Cluster Identification in Nearest-Neighbor Graphs

Maier, M., Hein, M., von Luxburg, U.

(163), Max-Planck-Institute for Biological Cybernetics, Tübingen, Germany, May 2007 (techreport)

Abstract
Assume we are given a sample of points from some underlying distribution which contains several distinct clusters. Our goal is to construct a neighborhood graph on the sample points such that clusters are ``identified‘‘: that is, the subgraph induced by points from the same cluster is connected, while subgraphs corresponding to different clusters are not connected to each other. We derive bounds on the probability that cluster identification is successful, and use them to predict ``optimal‘‘ values of k for the mutual and symmetric k-nearest-neighbor graphs. We point out different properties of the mutual and symmetric nearest-neighbor graphs related to the cluster identification problem.

ei

PDF [BibTex]

PDF [BibTex]


no image
Dirichlet Mixtures of Bayesian Linear Gaussian State-Space Models: a Variational Approach

Chiappa, S., Barber, D.

(161), Max Planck Institute for Biological Cybernetics, Tübingen, Germany, March 2007 (techreport)

Abstract
We describe two related models to cluster multidimensional time-series under the assumption of an underlying linear Gaussian dynamical process. In the first model, times-series are assigned to the same cluster when they show global similarity in their dynamics, while in the second model times-series are assigned to the same cluster when they show simultaneous similarity. Both models are based on Dirichlet Mixtures of Bayesian Linear Gaussian State-Space Models in order to (semi) automatically determine an appropriate number of components in the mixture, and to additionally bias the components to a parsimonious parameterization. The resulting models are formally intractable and to deal with this we describe a deterministic approximation based on a novel implementation of Variational Bayes.

ei

PDF [BibTex]

PDF [BibTex]


no image
Automatic 3D Face Reconstruction from Single Images or Video

Breuer, P., Kim, K., Kienzle, W., Blanz, V., Schölkopf, B.

(160), Max Planck Institute for Biological Cybernetics, Tübingen, Germany, February 2007 (techreport)

Abstract
This paper presents a fully automated algorithm for reconstructing a textured 3D model of a face from a single photograph or a raw video stream. The algorithm is based on a combination of Support Vector Machines (SVMs) and a Morphable Model of 3D faces. After SVM face detection, individual facial features are detected using a novel regression-and classification-based approach, and probabilistically plausible configurations of features are selected to produce a list of candidates for several facial feature positions. In the next step, the configurations of feature points are evaluated using a novel criterion that is based on a Morphable Model and a combination of linear projections. Finally, the feature points initialize a model-fitting procedure of the Morphable Model. The result is a high-resolution 3D surface model.

ei

PDF [BibTex]

PDF [BibTex]


no image
Relative Entropy Policy Search

Peters, J.

CLMC Technical Report: TR-CLMC-2007-2, Computational Learning and Motor Control Lab, Los Angeles, CA, 2007, clmc (techreport)

Abstract
This technical report describes a cute idea of how to create new policy search approaches. It directly relates to the Natural Actor-Critic methods but allows the derivation of one shot solutions. Future work may include the application to interesting problems.

am ei

PDF link (url) [BibTex]

PDF link (url) [BibTex]


no image
Learning an Outlier-Robust Kalman Filter

Ting, J., Theodorou, E., Schaal, S.

CLMC Technical Report: TR-CLMC-2007-1, Los Angeles, CA, 2007, clmc (techreport)

Abstract
We introduce a modified Kalman filter that performs robust, real-time outlier detection, without the need for manual parameter tuning by the user. Systems that rely on high quality sensory data (for instance, robotic systems) can be sensitive to data containing outliers. The standard Kalman filter is not robust to outliers, and other variations of the Kalman filter have been proposed to overcome this issue. However, these methods may require manual parameter tuning, use of heuristics or complicated parameter estimation procedures. Our Kalman filter uses a weighted least squares-like approach by introducing weights for each data sample. A data sample with a smaller weight has a weaker contribution when estimating the current time step?s state. Using an incremental variational Expectation-Maximization framework, we learn the weights and system dynamics. We evaluate our Kalman filter algorithm on data from a robotic dog.

am

PDF [BibTex]

PDF [BibTex]

2005


no image
Popper, Falsification and the VC-dimension

Corfield, D., Schölkopf, B., Vapnik, V.

(145), Max Planck Institute for Biological Cybernetics, November 2005 (techreport)

ei

PDF [BibTex]

2005


PDF [BibTex]


no image
A Combinatorial View of Graph Laplacians

Huang, J.

(144), Max Planck Institute for Biological Cybernetics, Tübingen, Germany, August 2005 (techreport)

Abstract
Discussions about different graph Laplacian, mainly normalized and unnormalized versions of graph Laplacian, have been ardent with respect to various methods in clustering and graph based semi-supervised learning. Previous research on graph Laplacians investigated their convergence properties to Laplacian operators on continuous manifolds. There is still no strong proof on convergence for the normalized Laplacian. In this paper, we analyze different variants of graph Laplacians directly from the ways solving the original graph partitioning problem. The graph partitioning problem is a well-known combinatorial NP hard optimization problem. The spectral solutions provide evidence that normalized Laplacian encodes more reasonable considerations for graph partitioning. We also provide some examples to show their differences.

ei

[BibTex]

[BibTex]


no image
Beyond Pairwise Classification and Clustering Using Hypergraphs

Zhou, D., Huang, J., Schölkopf, B.

(143), Max Planck Institute for Biological Cybernetics, August 2005 (techreport)

Abstract
In many applications, relationships among objects of interest are more complex than pairwise. Simply approximating complex relationships as pairwise ones can lead to loss of information. An alternative for these applications is to analyze complex relationships among data directly, without the need to first represent the complex relationships into pairwise ones. A natural way to describe complex relationships is to use hypergraphs. A hypergraph is a graph in which edges can connect more than two vertices. Thus we consider learning from a hypergraph, and develop a general framework which is applicable to classification and clustering for complex relational data. We have applied our framework to real-world web classification problems and obtained encouraging results.

ei

PDF [BibTex]

PDF [BibTex]


no image
Measuring Statistical Dependence with Hilbert-Schmidt Norms

Gretton, A., Bousquet, O., Smola, A., Schölkopf, B.

(140), Max Planck Institute for Biological Cybernetics, Tübingen, Germany, June 2005 (techreport)

Abstract
We propose an independence criterion based on the eigenspectrum of covariance operators in reproducing kernel Hilbert spaces (RKHSs), consisting of an empirical estimate of the Hilbert-Schmidt norm of the cross-covariance operator (we term this a Hilbert-Schmidt Independence Criterion, or HSIC). This approach has several advantages, compared with previous kernel-based independence criteria. First, the empirical estimate is simpler than any other kernel dependence test, and requires no user-defined regularisation. Second, there is a clearly defined population quantity which the empirical estimate approaches in the large sample limit, with exponential convergence guaranteed between the two: this ensures that independence tests based on HSIC do not suffer from slow learning rates. Finally, we show in the context of independent component analysis (ICA) that the performance of HSIC is competitive with that of previously published kernel-based criteria, and of other recently published ICA methods.

ei

PDF [BibTex]

PDF [BibTex]


no image
Approximate Inference for Robust Gaussian Process Regression

Kuss, M., Pfingsten, T., Csato, L., Rasmussen, C.

(136), Max Planck Institute for Biological Cybernetics, Tübingen, Germany, 2005 (techreport)

Abstract
Gaussian process (GP) priors have been successfully used in non-parametric Bayesian regression and classification models. Inference can be performed analytically only for the regression model with Gaussian noise. For all other likelihood models inference is intractable and various approximation techniques have been proposed. In recent years expectation-propagation (EP) has been developed as a general method for approximate inference. This article provides a general summary of how expectation-propagation can be used for approximate inference in Gaussian process models. Furthermore we present a case study describing its implementation for a new robust variant of Gaussian process regression. To gain further insights into the quality of the EP approximation we present experiments in which we compare to results obtained by Markov chain Monte Carlo (MCMC) sampling.

ei

PDF [BibTex]

PDF [BibTex]


no image
Maximum-Margin Feature Combination for Detection and Categorization

BakIr, G., Wu, M., Eichhorn, J.

Max Planck Institute for Biological Cybernetics, Tübingen, Germany, 2005 (techreport)

Abstract
In this paper we are concerned with the optimal combination of features of possibly different types for detection and estimation tasks in machine vision. We propose to combine features such that the resulting classifier maximizes the margin between classes. In contrast to existing approaches which are non-convex and/or generative we propose to use a discriminative model leading to convex problem formulation and complexity control. Furthermore we assert that decision functions should not compare apples and oranges by comparing features of different types directly. Instead we propose to combine different similarity measures for each different feature type. Furthermore we argue that the question: ”Which feature type is more discriminative for task X?” is ill-posed and show empirically that the answer to this question might depend on the complexity of the decision function.

ei

PDF [BibTex]

PDF [BibTex]


no image
Towards a Statistical Theory of Clustering. Presented at the PASCAL workshop on clustering, London

von Luxburg, U., Ben-David, S.

Presented at the PASCAL workshop on clustering, London, 2005 (techreport)

Abstract
The goal of this paper is to discuss statistical aspects of clustering in a framework where the data to be clustered has been sampled from some unknown probability distribution. Firstly, the clustering of the data set should reveal some structure of the underlying data rather than model artifacts due to the random sampling process. Secondly, the more sample points we have, the more reliable the clustering should be. We discuss which methods can and cannot be used to tackle those problems. In particular we argue that generalization bounds as they are used in statistical learning theory of classification are unsuitable in a general clustering framework. We suggest that the main replacements of generalization bounds should be convergence proofs and stability considerations. This paper should be considered as a road map paper which identifies important questions and potentially fruitful directions for future research about statistical clustering. We do not attempt to present a complete statistical theory of clustering.

ei

PDF [BibTex]

PDF [BibTex]


no image
Approximate Bayesian Inference for Psychometric Functions using MCMC Sampling

Kuss, M., Jäkel, F., Wichmann, F.

(135), Max Planck Institute for Biological Cybernetics, Tübingen, Germany, 2005 (techreport)

Abstract
In psychophysical studies the psychometric function is used to model the relation between the physical stimulus intensity and the observer's ability to detect or discriminate between stimuli of different intensities. In this report we propose the use of Bayesian inference to extract the information contained in experimental data estimate the parameters of psychometric functions. Since Bayesian inference cannot be performed analytically we describe how a Markov chain Monte Carlo method can be used to generate samples from the posterior distribution over parameters. These samples are used to estimate Bayesian confidence intervals and other characteristics of the posterior distribution. In addition we discuss the parameterisation of psychometric functions and the role of prior distributions in the analysis. The proposed approach is exemplified using artificially generate d data and in a case study for real experimental data. Furthermore, we compare our approach with traditional methods based on maximum-likelihood parameter estimation combined with bootstrap techniques for confidence interval estimation. The appendix provides a description of an implementation for the R environment for statistical computing and provides the code for reproducing the results discussed in the experiment section.

ei

PDF [BibTex]

PDF [BibTex]


no image
Linear and Nonlinear Estimation models applied to Hemodynamic Model

Theodorou, E.

Technical Report-2005-1, Computational Action and Vision Lab University of Minnesota, 2005, clmc (techreport)

Abstract
The relation between BOLD signal and neural activity is still poorly understood. The Gaussian Linear Model known as GLM is broadly used in many fMRI data analysis for recovering the underlying neural activity. Although GLM has been proved to be a really useful tool for analyzing fMRI data it can not be used for describing the complex biophysical process of neural metabolism. In this technical report we make use of a system of Stochastic Differential Equations that is based on Buxton model [1] for describing the underlying computational principles of hemodynamic process. Based on this SDE we built a Kalman Filter estimator so as to estimate the induced neural signal as well as the blood inflow under physiologic and sensor noise. The performance of Kalman Filter estimator is investigated under different physiologic noise characteristics and measurement frequencies.

am

PDF [BibTex]

PDF [BibTex]

2004


no image
Joint Kernel Maps

Weston, J., Schölkopf, B., Bousquet, O., Mann, .., Noble, W.

(131), Max-Planck-Institute for Biological Cybernetics, Tübingen, November 2004 (techreport)

ei

PDF [BibTex]

2004


PDF [BibTex]


no image
Semi-Supervised Induction

Yu, K., Tresp, V., Zhou, D.

(141), Max Planck Institute for Biological Cybernetics, Tuebingen, Germany, August 2004 (techreport)

Abstract
Considerable progress was recently achieved on semi-supervised learning, which differs from the traditional supervised learning by additionally exploring the information of the unlabelled examples. However, a disadvantage of many existing methods is that it does not generalize to unseen inputs. This paper investigates learning methods that effectively make use of both labelled and unlabelled data to build predictive functions, which are defined on not just the seen inputs but the whole space. As a nice property, the proposed method allows effcient training and can easily handle new test points. We validate the method based on both toy data and real world data sets.

ei

PDF PDF [BibTex]

PDF PDF [BibTex]


no image
Object categorization with SVM: kernels for local features

Eichhorn, J., Chapelle, O.

(137), Max Planck Institute for Biological Cybernetics, Tübingen, Germany, July 2004 (techreport)

Abstract
In this paper, we propose to combine an efficient image representation based on local descriptors with a Support Vector Machine classifier in order to perform object categorization. For this purpose, we apply kernels defined on sets of vectors. After testing different combinations of kernel / local descriptors, we have been able to identify a very performant one.

ei

PDF [BibTex]

PDF [BibTex]


no image
Hilbertian Metrics and Positive Definite Kernels on Probability Measures

Hein, M., Bousquet, O.

(126), Max Planck Institute for Biological Cybernetics, Tübingen, Germany, July 2004 (techreport)

Abstract
We investigate the problem of defining Hilbertian metrics resp. positive definite kernels on probability measures, continuing previous work. This type of kernels has shown very good results in text classification and has a wide range of possible applications. In this paper we extend the two-parameter family of Hilbertian metrics of Topsoe such that it now includes all commonly used Hilbertian metrics on probability measures. This allows us to do model selection among these metrics in an elegant and unified way. Second we investigate further our approach to incorporate similarity information of the probability space into the kernel. The analysis provides a better understanding of these kernels and gives in some cases a more efficient way to compute them. Finally we compare all proposed kernels in two text and one image classification problem.

ei

PDF [BibTex]

PDF [BibTex]


no image
Kernels, Associated Structures and Generalizations

Hein, M., Bousquet, O.

(127), Max Planck Institute for Biological Cybernetics, Tübingen, Germany, July 2004 (techreport)

Abstract
This paper gives a survey of results in the mathematical literature on positive definite kernels and their associated structures. We concentrate on properties which seem potentially relevant for Machine Learning and try to clarify some results that have been misused in the literature. Moreover we consider different lines of generalizations of positive definite kernels. Namely we deal with operator-valued kernels and present the general framework of Hilbertian subspaces of Schwartz which we use to introduce kernels which are distributions. Finally indefinite kernels and their associated reproducing kernel spaces are considered.

ei

PDF [BibTex]

PDF [BibTex]


no image
Kamerakalibrierung und Tiefenschätzung: Ein Vergleich von klassischer Bündelblockausgleichung und statistischen Lernalgorithmen

Sinz, FH.

Wilhelm-Schickard-Institut für Informatik, Universität Tübingen, Tübingen, Germany, March 2004 (techreport)

Abstract
Die Arbeit verleicht zwei Herangehensweisen an das Problem der Sch{\"a}tzung der r{\"a}umliche Position eines Punktes aus den Bildkoordinaten in zwei verschiedenen Kameras. Die klassische Methode der B{\"u}ndelblockausgleichung modelliert zwei Einzelkameras und sch{\"a}tzt deren {\"a}ußere und innere Orientierung mit einer iterativen Kalibrationsmethode, deren Konvergenz sehr stark von guten Startwerten abh{\"a}ngt. Die Tiefensch{\"a}tzung eines Punkts geschieht durch die Invertierung von drei der insgesamt vier Projektionsgleichungen der Einzalkameramodelle. Die zweite Methode benutzt Kernel Ridge Regression und Support Vector Regression, um direkt eine Abbildung von den Bild- auf die Raumkoordinaten zu lernen. Die Resultate zeigen, daß der Ansatz mit maschinellem Lernen, neben einer erheblichen Vereinfachung des Kalibrationsprozesses, zu h{\"o}heren Positionsgenaugikeiten f{\"u}hren kann.

ei

PDF [BibTex]

PDF [BibTex]


no image
Multivariate Regression with Stiefel Constraints

Bakir, G., Gretton, A., Franz, M., Schölkopf, B.

(128), MPI for Biological Cybernetics, Spemannstr 38, 72076, Tuebingen, 2004 (techreport)

Abstract
We introduce a new framework for regression between multi-dimensional spaces. Standard methods for solving this problem typically reduce the problem to one-dimensional regression by choosing features in the input and/or output spaces. These methods, which include PLS (partial least squares), KDE (kernel dependency estimation), and PCR (principal component regression), select features based on different a-priori judgments as to their relevance. Moreover, loss function and constraints are chosen not primarily on statistical grounds, but to simplify the resulting optimisation. By contrast, in our approach the feature construction and the regression estimation are performed jointly, directly minimizing a loss function that we specify, subject to a rank constraint. A major advantage of this approach is that the loss is no longer chosen according to the algorithmic requirements, but can be tailored to the characteristics of the task at hand; the features will then be optimal with respect to this objective. Our approach also allows for the possibility of using a regularizer in the optimization. Finally, by processing the observations sequentially, our algorithm is able to work on large scale problems.

ei

PDF [BibTex]

PDF [BibTex]


no image
Learning from Labeled and Unlabeled Data Using Random Walks

Zhou, D., Schölkopf, B.

Max Planck Institute for Biological Cybernetics, 2004 (techreport)

Abstract
We consider the general problem of learning from labeled and unlabeled data. Given a set of points, some of them are labeled, and the remaining points are unlabeled. The goal is to predict the labels of the unlabeled points. Any supervised learning algorithm can be applied to this problem, for instance, Support Vector Machines (SVMs). The problem of our interest is if we can implement a classifier which uses the unlabeled data information in some way and has higher accuracy than the classifiers which use the labeled data only. Recently we proposed a simple algorithm, which can substantially benefit from large amounts of unlabeled data and demonstrates clear superiority to supervised learning methods. In this paper we further investigate the algorithm using random walks and spectral graph theory, which shed light on the key steps in this algorithm.

ei

PDF PostScript [BibTex]

PDF PostScript [BibTex]


no image
Behaviour and Convergence of the Constrained Covariance

Gretton, A., Smola, A., Bousquet, O., Herbrich, R., Schölkopf, B., Logothetis, N.

(130), MPI for Biological Cybernetics, 2004 (techreport)

Abstract
We discuss reproducing kernel Hilbert space (RKHS)-based measures of statistical dependence, with emphasis on constrained covariance (COCO), a novel criterion to test dependence of random variables. We show that COCO is a test for independence if and only if the associated RKHSs are universal. That said, no independence test exists that can distinguish dependent and independent random variables in all circumstances. Dependent random variables can result in a COCO which is arbitrarily close to zero when the source densities are highly non-smooth, which can make dependence hard to detect empirically. All current kernel-based independence tests share this behaviour. Finally, we demonstrate exponential convergence between the population and empirical COCO, which implies that COCO does not suffer from slow learning rates when used as a dependence test.

ei

PDF [BibTex]

PDF [BibTex]


no image
Confidence Sets for Ratios: A Purely Geometric Approach To Fieller’s Theorem

von Luxburg, U., Franz, V.

(133), Max Planck Institute for Biological Cybernetics, 2004 (techreport)

Abstract
We present a simple, geometric method to construct Fieller's exact confidence sets for ratios of jointly normally distributed random variables. Contrary to previous geometric approaches in the literature, our method is valid in the general case where both sample mean and covariance are unknown. Moreover, not only the construction but also its proof are purely geometric and elementary, thus giving intuition into the nature of the confidence sets.

ei

PDF [BibTex]

PDF [BibTex]


no image
Transductive Inference with Graphs

Zhou, D., Schölkopf, B.

Max Planck Institute for Biological Cybernetics, 2004, See the improved version Regularization on Discrete Spaces. (techreport)

Abstract
We propose a general regularization framework for transductive inference. The given data are thought of as a graph, where the edges encode the pairwise relationships among data. We develop discrete analysis and geometry on graphs, and then naturally adapt the classical regularization in the continuous case to the graph situation. A new and effective algorithm is derived from this general framework, as well as an approach we developed before.

ei

[BibTex]

[BibTex]